parameters?

[HaruoHosoya]

I have difficulty of reproducing the results shown from README. I ran ff_training.py, get_rpr.py, and invert.py consecutively using the same versions of python/pytorch, but I got only unrecognizable noise images as reconstruction. I ran it several times but the results are similar. Did you give additional command-line arguments to these scripts?

[ukky17]

@HaruoHosoya Thank you for the comment. I didn't give any additional arguments; I used the default parameters given in ArgumentParser and was able to generate naturalistic images.

I don't know why the program didn't work in your case, but the loss scores might be helpful to adjust the parameters, especially lambdas. For your reference, below is the output when I previously run python invert.py (with the default parameters).

test resp shape: (10000, 256, 8, 8)
test image shape: (10000, 3, 32, 32)
train resp shape: (50000, 256, 8, 8)
train image shape: (50000, 3, 32, 32)
Epoch 0, Time 238.12
  Train: 0.012796, 0.547893, 0.069729, 0.014041, 0.300302
  Test:  0.010291, 0.493790, 0.058010, 0.011365, 0.274226
Epoch 1, Time 235.47
  Train: 0.010145, 0.539667, 0.056430, 0.011249, 0.301747
  Test:  0.009722, 0.668325, 0.052899, 0.010919, 0.327969
TEpoch 2, Time 234.49
  Train: 0.009468, 0.529180, 0.052715, 0.010525, 0.307933
  Test:  0.009044, 0.522657, 0.051326, 0.010080, 0.256956
Epoch 3, Time 234.92
  Train: 0.009080, 0.524205, 0.050421, 0.010109, 0.311537
  Test:  0.009157, 0.442447, 0.050725, 0.010106, 0.291684
Epoch 4, Time 234.66
  Train: 0.008813, 0.519279, 0.048985, 0.009822, 0.315642
  Test:  0.008685, 0.483414, 0.048966, 0.009658, 0.260552
...

[HaruoHosoya]

@ukky17 Thanks for reply. I show below two logs, the one using the default parameters and the other using adjusted lambda_* parameters to compensate the difference between yours and mine (the ratio loss_g/loss_d for training now looks similar to yours). The resulting images are similar and bad. I notice that, in my case, loss_d values for training and test are very different, while yours are similar. Also, in my case, loss_g doesn't decrease at all, while yours does. Any idea?

1. using the default parameters

test resp shape: (10000, 256, 8, 8)
test image shape: (10000, 3, 32, 32)
train resp shape: (50000, 256, 8, 8)
train image shape: (50000, 3, 32, 32)
Epoch 0, Time 69.04
Train: 0.062342, 0.920429, 0.180257, 0.065065, 0.070150
Test: 0.061357, 0.916260, 0.176226, 0.064036, 0.004469
Epoch 1, Time 66.64
Train: 0.062115, 0.956813, 0.176591, 0.064838, 0.037477
Test: 0.061169, 1.052739, 0.173898, 0.063961, 0.003879
Epoch 2, Time 67.17
Train: 0.062082, 0.965750, 0.175825, 0.064806, 0.030368
Test: 0.060777, 0.863741, 0.176078, 0.063402, 0.008579
Epoch 3, Time 67.46
Train: 0.062075, 0.966654, 0.176177, 0.064803, 0.034042
Test: 0.060959, 0.954981, 0.174026, 0.063654, 0.005351
Epoch 4, Time 66.99
Train: 0.062052, 0.970124, 0.174856, 0.064771, 0.028622
Test: 0.061040, 1.082508, 0.173847, 0.063861, 0.004120
Epoch 5, Time 66.90
Train: 0.062035, 0.973467, 0.174632, 0.064755, 0.020890
Test: 0.061127, 1.007541, 0.174084, 0.063875, 0.001082
Epoch 6, Time 67.45
Train: 0.062045, 0.968793, 0.175321, 0.064767, 0.036514
Test: 0.061175, 1.041520, 0.173190, 0.063948, 0.004075
Epoch 7, Time 67.51
Train: 0.062033, 0.973549, 0.175029, 0.064757, 0.027554
Test: 0.060844, 0.989257, 0.174212, 0.063576, 0.005379
Epoch 8, Time 67.42
Train: 0.062007, 0.978004, 0.174833, 0.064734, 0.021346
Test: 0.061076, 1.053426, 0.172262, 0.063852, 0.001646
...

2. using adjusted parameters

python invert.py --lambda_img=0.02 --lambda_adv=0.00005 --lambda_feat=0.0003
test resp shape: (10000, 256, 8, 8)
test image shape: (10000, 3, 32, 32)
train resp shape: (50000, 256, 8, 8)
train image shape: (50000, 3, 32, 32)
Epoch 0, Time 70.61
Train: 0.062621, 0.920313, 0.179995, 0.001352, 0.096919
Test: 0.061623, 1.006139, 0.174405, 0.001335, 0.019429
Epoch 1, Time 70.51
Train: 0.062315, 0.963888, 0.175787, 0.001347, 0.036341
Test: 0.060994, 1.037385, 0.173458, 0.001324, 0.008299
Epoch 2, Time 70.41
Train: 0.062280, 0.955017, 0.174686, 0.001346, 0.040562
Test: 0.061169, 0.975571, 0.175972, 0.001325, 0.007981
Epoch 3, Time 68.71
Train: 0.062242, 0.978318, 0.174623, 0.001346, 0.041942
Test: 0.061127, 0.977562, 0.173377, 0.001323, 0.002352
Epoch 4, Time 68.49
Train: 0.062210, 0.970575, 0.173272, 0.001345, 0.027981
Test: 0.060934, 1.018060, 0.173111, 0.001322, 0.003301
Epoch 5, Time 68.62
Train: 0.062224, 0.972491, 0.173549, 0.001345, 0.031365
Test: 0.060962, 1.112787, 0.171626, 0.001326, 0.005110
Epoch 6, Time 68.07
Train: 0.062190, 0.972276, 0.173274, 0.001344, 0.026025
Test: 0.060579, 0.934411, 0.172045, 0.001310, 0.008814
Epoch 7, Time 68.46
Train: 0.062165, 0.969345, 0.171951, 0.001343, 0.029999
Test: 0.061101, 0.908433, 0.171522, 0.001319, 0.005193
Epoch 8, Time 68.78
Train: 0.062248, 0.969179, 0.172746, 0.001345, 0.038489
Test: 0.061236, 1.065461, 0.171845, 0.001330, 0.004004
Epoch 9, Time 67.84
Train: 0.062204, 0.973345, 0.172559, 0.001345, 0.033740
Test: 0.061222, 1.016806, 0.173432, 0.001327, 0.006019
Epoch 10, Time 68.05
Train: 0.062219, 0.969613, 0.172944, 0.001345, 0.034513
Test: 0.061145, 1.009146, 0.170403, 0.001324, 0.002087
Epoch 11, Time 67.76
Train: 0.062147, 0.968081, 0.172071, 0.001343, 0.028754
Test: 0.061019, 0.990229, 0.172044, 0.001322, 0.002990
Epoch 12, Time 67.92
Train: 0.062125, 0.974244, 0.172498, 0.001343, 0.030615
Test: 0.061018, 1.023234, 0.171053, 0.001323, 0.005527
Epoch 13, Time 68.24
Train: 0.062164, 0.972657, 0.173308, 0.001344, 0.031809
Test: 0.060929, 0.911533, 0.172117, 0.001316, 0.006186
..

[ukky17]

@HaruoHosoya In your both cases, loss_d is almost zero from the beginning, which means the generated noisy images are too easy to be classified as fake. I guess you need to increment lambda_adv (e.g., x5, x10, x100, ...) over the other lambdas to create more naturalistic images that are more challenging to the descriminator.